Installation

gplearn-CUDA requires a recent version of scikit-learn (which requires numpy and scipy). So first you will need to follow their installation instructions to get the dependencies.

Python support currently begins at Python 3.11, matching the requirement imposed by scikit-learn>=1.8.0.

Option 1: installation using pip

Now that you have scikit-learn installed, you can install gplearn-CUDA using pip:

pip install gplearn-CUDA

The base package does not install CuPy or CUDA runtime libraries. That is intentional so pip install gplearn-CUDA stays lightweight and works on machines without NVIDIA CUDA support.

To enable GPU/CUDA acceleration, install one of the optional CUDA extras. gplearn-CUDA is compatible with CUDA 11.2 through 13.x. The extras use CuPy’s [ctk] bundle so common CUDA runtime libraries such as cuBLAS, cuRAND, cuSPARSE, cuSOLVER, cuFFT, cudart and NVRTC are installed alongside the matching CuPy wheel:

pip install "gplearn-CUDA[cuda13]"

The cuda extra is a convenience alias for the current recommended CUDA 13.x path:

pip install "gplearn-CUDA[cuda]"

If you need an explicit CUDA 12.x install path, use:

pip install "gplearn-CUDA[cuda12]"

The easiest way to install the correct GPU drivers and libraries is via Conda:

conda install -c conda-forge cupy

If you prefer using pip manually, you must install the specific CuPy wheel that matches your system’s CUDA version, then install gplearn-CUDA itself:

  • For CUDA 13.x: pip install "cupy-cuda13x[ctk]"

  • For CUDA 12.x: pip install "cupy-cuda12x[ctk]"

  • For CUDA 11.x: pip install cupy-cuda11x

Then:

pip install gplearn-CUDA

If you want to install CuPy separately while still bundling the CUDA toolkit runtime libraries, use:

pip install "cupy-cuda13x[ctk]"

Or if you wish to install to the home directory:

pip install --user gplearn-CUDA

For the latest development version, first get the source from github:

git clone https://github.com/LGA-Personal/gplearn-CUDA.git

Then navigate into the local gplearn-CUDA directory and simply run:

pip install .

or:

pip install --user .

and you’re done!

Option 2: installation using conda

In case you want to install gplearn-CUDA in an anaconda environment, you can run:

conda install -c conda-forge gplearn-CUDA

and you’re done!

Notes

  • Verified CPU support in this repository covers Python 3.11, 3.12, 3.13 and 3.14 on Windows.

  • Verified CUDA support in this repository covers Python 3.12 and 3.14 on Windows.

  • On some Windows Python 3.14 environments, CuPy can import and compile CUDA kernels while optional BLAS backends such as cuBLAS still fail to load. In that case, SymbolicTransformer(device='cuda') falls back to a NumPy correlation step and emits a warning instead of aborting.